Pytorch implementation for GuiDG

PyTorch implementation of Generalizing Vision-Language Models with Dedicated Prompt Guidance (AAAI'26). The following guidance runs GuiDG on ImageNet-DG. The code supports other datasets reported in paper with similar usage.

• Python==3.12, Pytorch==2.4.1
• Clone CoOp and prepare environments as instructed (including Dassl, CLIP, etc.).
• Clone DomainBed and prepare environments as instructed.
• Important: Comment Line 5 in Dassl.pytorch-master/dassl/data/datasets/__init__.py (otherwise there might not be outputs in the logs):

# from .dg import *

• Important: Add the following code between Line 223-224 of CLIP/clip/models.py (check CLIPood):

x = x.type(self.conv1.weight.dtype)

• Move imagenet_dg.py, officehome.py, terra_incognita.py, pacs.py, domainnet.py to CoOp/dataset/.

• Download ImageNet-A, ImageNet-R, ImageNet-V2, ImageNet-Sketch, ImageNet as instructed in CoOp.
• Soft link the downloaded datasets to DomainBed/domainbed/data/ImageNet/ to obtain the directory structure as follows (a for ImageNet-A, i for ImageNet, r for ImageNet-R, s for ImageNet-Sketch, v2 for ImageNet-V2):

ImageNet/
|-- a/
    |-- n01498041/
    |-- ...... (200 folders)
    |-- classnames.txt
|-- i/
    |-- train/
        |-- n01440764/
        |-- ...... (1000 folders)
        |-- classnames.txt
    |-- val/
        |-- n01440764/
        |-- ...... (1000 folders)
        |-- classnames.txt
|-- r/
    |-- n01443537/
    ...... (200 folders)
    |-- classnames.txt
|-- s/
    |-- n01440764/
    |-- ...... (1000 folders)
    |-- classnames.txt
|-- v2/
    |-- n01440764/
    |-- ...... (1000 folders)
    |-- classnames.txt
|-- classnames.txt

• Train your own domain experts first (requires >= 24G GPU):

bash scripts/ImageNetDG_Step1.sh

• Then fine-tune CLIP with dedicated prompt guidance:

bash scripts/ImageNetDG_Step2.sh

• Check log/ for outputs.